← Back to work

Havenly

Pilot in buildWatercolor bloomGitHub ↗
ProblemPeople bounce off wellness apps the moment they feel judged. Could a daily companion feel safe enough to return to?
My roleSole designer & builder — product design through to a live 3-service backend
OutcomePilot running end-to-end: 3-service chain live, 120+ automated tests, 3 dangerous defects caught in review before anyone could hit them.
3
services live, end-to-end
120+
automated tests
3
dangerous defects caught pre-ship
THE BET

A pilot in build, not a shipped product

Havenly is a daily AI wellness companion — a calm place to check in, not a metrics dashboard. It is a pilot in build, not a shipped product. What makes it portfolio evidence is that I designed and built it: the interaction model, the safety architecture, and the running backend.

01 · SAFETY

Designing for safety first

A wellness companion that talks back has to fail safe. The crisis-safety path is a first-class flow, not an afterthought — and the adversarial review leg of the pipeline caught a defect where a specific phrasing could route a user past the crisis check. The ordering was reworked before it ever reached a person.

02 · REVIEW

What the agents caught

Three dangerous defects surfaced in adversarial and independent review: the crisis-safety routing above, a data-erasure ordering bug that could delete a journal before confirming intent, and an A/B design that would have produced uninterpretable results. None were mine to be proud of — catching them is.

03 · STATE

Where it stands

Live end-to-end across three services with 120+ tests. The design system, the companion's voice, and the safety rails are the deliverable; the metrics come after the pilot. Honest label, honest state.

See how I ship with agentsBack to work